Fourier Neural Operator (FNO)

The Fourier Neural Operator (FNO) is a neural operator architecture introduced by Zongyi Li, Nikola Kovachki, Kamyar Azizzadeneshrani, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar in 2020 (arXiv 2010.08895; ICLR 2020 Notable — Top 5%). FNO parameterizes the learned integral kernel of a neural operator block in Fourier space rather than in physical space, replacing the costly kernel-integration step with a Fourier transform, a learned truncation-and-multiplication of the lower-frequency modes, and an inverse Fourier transform.

The motivation is that for many parametric PDEs on bounded domains — Burgers, Darcy flow, Navier-Stokes, weather dynamics — the relevant features of the solution live in the lower-frequency Fourier modes. Truncating to the first K modes (typically K = 16 to 32) gives a tractable architecture that is fast to train and evaluate while retaining accuracy. FNO was the first architecture to demonstrate that a learned operator could match or exceed the accuracy of finite-difference PDE solvers on standard benchmarks at dramatically lower inference cost, often running 1000× to 10000× faster than the solvers they replace once trained.

FNO is the most widely deployed neural operator variant in scientific machine learning. Variants and descendants include Geo-FNO (irregular geometries), U-FNO (U-Net-like multi-resolution stack), Factorized FNO (factorized spectral modes), and FNO-based foundation models for weather (FourCastNet, Pangu-Weather, GraphCast's operator components) and industrial CFD. It is the canonical workhorse of operator learning on structured grids, complementing Graph Neural Operator (irregular) and DeepONet (branch/trunk factorization).